Claude Fable 5: The Autonomous Co-Worker Above Opus
Most people are still thinking of AI models as faster chatbots. Claude Fable 5 is not that, and if you want to understand the tier sitting above the Opus class, this is for you. Fable 5 was recently released to the public as the first of its class, tuned for long, complex, multi-stage work that a quick chatbot reply could never handle. The video above is a full breakdown, and below is the written version of what Fable 5 is, what it proved it can do, and the guardrails and cost that came with it, all as covered at the time of recording.
A tier above Opus
Fable 5 sits above the Opus class, shares its underlying model with Metis 5, and adds extra safeguards.
Fable 5 is the first model of a new class that sits above the Opus tier in raw capability. Under the hood it is essentially the same as Metis 5, but it is unique in that it ships with additional safeguards built in. It is tuned for long, complex, multi-stage tasks that were not really possible before.
On the benchmarking shared at recording, it landed roughly 10 percent above the most recent Opus 4.8 and passed model tests around 90 percent accurate. Its focus is the hard end of the scale: the harder the task, the more it can do and the more accurate it becomes.
It works for days, and checks itself
Fable 5 can run multi-day tasks autonomously with self-checking that guards its own accuracy.
The headline shift is duration. This model can work for multiple days and do the work on its own, and it has self-checking capabilities that let it guard and correct itself, which feeds directly back into accuracy.
The highlight that stuck with me from the recording was roughly 50 million lines of Ruby migrated in a single day, work that would have taken many months. On token efficiency it did the same tasks with fewer tokens, which is a meaningful cost reduction. What I found notable is that the compression of time, not just the raw quality, is what changes how you would actually use it.
What it proved it can do
Fable 5 aced deep finance and trading analysis, beat human-level legal redlines, and read dense diagrams buried in PDFs.
The range covered at recording was wide. On finance it handled deep research across numbers, charts, tables, and documents, and on trading it aced evaluations on root cause and expected-value reasoning. If you have ever built any of that analysis by hand, you know how much time that represents. On legal it handled redline edits better than a human would, which anyone who has slogged through a contract knows is no small thing.
It also read dense diagrams and tables buried inside PDFs, and checked code against screenshots for accuracy, an area older models struggled with. Think about how often you have needed exactly that and could not trust it. On memory it showed a 3x payoff over Opus 4.8 in a Slay the Spire test, and it even beat Pokemon from screenshots.
The guardrails that shipped with it
Fable 5 added specific guardrails for cybersecurity, biology, chemistry, and model distillation, backed by heavy red-team testing.
The safeguards are a real part of the story. Researchers vetted the model, and the guardrails focus on cybersecurity, biology and chemistry, and model distillation. At recording it had survived over 1,000 hours of bug-bounty testing with no universal jailbreak found, and it carried a new 30-day data-retention policy on this tier.
The capability claims came with accuracy claims too: a roughly 10x faster drug-design step in testing, and scientists hypothesizing at around 80 percent accuracy on blind tests. The point is that the guardrails and the accuracy shipped together, not as an afterthought.
Cost and where to get it
At recording Fable 5 ran about 10 dollars per million input tokens and 50 per million output, available through Claude apps, the API, and major clouds.
On price, the figures shared were roughly 10 dollars per million input tokens and 50 dollars per million output tokens, about double Opus 4.8 and less than half the price of the Mythos preview. You could reach it through the Claude apps, the API, AWS Bedrock, and Google Cloud, with free access on Pro, Max, Team, and Enterprise through a launch window before it shifted to usage credits.
Early partners included Cursor and GitHub, and I built my read of it around that launch context. Treat every one of these numbers as a snapshot from the recording, since pricing and access on this tier move fast.
The takeaway
Fable 5 is not a faster chatbot, it is an autonomous co-worker. It sits above the Opus class, shares its model with Metis 5, adds real safeguards, and is tuned for days-long, self-checking, agentic work, landing about 10 percent above Opus 4.8 on the benchmarks shared at recording. It migrated tens of millions of lines of code in a day, aced finance, trading, and legal evaluations, and shipped guardrails for cybersecurity, biology, and chemistry alongside its capability gains. The mindset shift is away from prompting and toward handing off long work and reviewing what came back.
Watch the full breakdown in my video on Claude Fable 5, and I go through the capability highlights in the video. I also cover the guardrails and pricing in the video. Here is my question for the comments: what days-long task would you actually trust a model like this with? Subscribe for more deep dives as these models ship.
